U tweetove putem weba ili aplikacija drugih proizvođača možete dodati podatke o lokaciji, kao što su grad ili točna lokacija. Povijest lokacija tweetova uvijek možete izbrisati. Saznajte više
Data augmentation is even more crucial for detection. We present AutoAugment for object detection, achieving SOTA on COCO validation set (50.7 mAP). Policy transfers to different models & datasets. Paper: https://arxiv.org/abs/1906.11172 , Code: https://github.com/tensorflow/tpu/tree/master/models/official/detection …, details in thread.pic.twitter.com/XZuyMjrxx8
Thanks to AutoAugment, our model does not need special tricks previously used by detection models, and evaluates at a single spatial scale to achieve SOTA. The computational overhead at training time is negligible, and the computational overhead at test time is zero.pic.twitter.com/4jmzmc93SM
An important decision was about what to do with bounding boxes after rotation&shearing. We make the box large enough to contain all the pixels from pre-rotation, so the boxes aren't as tight as they can be. Despite this, rotation is very beneficial, which deserves further study.pic.twitter.com/wwyU4w2rDk
Detection allows more complex augmentations than classification. For example, we added ops that only act on bounding box contents. In these example images, only the bounding box contents are sheared.pic.twitter.com/qudYKRViIZ
Another interesting result is that AutoAugment helps more on harder tasks: detection of small objects and at larger IOU thresholds. Collaboration with @barret_zoph, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and @quocleixpic.twitter.com/8EceerCdtp
Did you try rotating the BB only (i.e. sub-policy 2 but rotation instead of shear)?
Yes, that was in the search space (we had the bounding-box-only version of all operations).pic.twitter.com/zs4ETWdtC1
Loved this! Do you know of pytorch/pure python implementations of your transforms? I'm working on a medical object detection model with pytorch and albumentations and would love to use your learnings!
For classification, here is a pytorch implementation (not by us): https://github.com/DeepVoltaire/AutoAugment … and here is our implementation using Python Image Library: https://github.com/tensorflow/models/blob/master/research/autoaugment/augmentation_transforms.py … These ops could also be used for detection, but don't know of a detection specific non-tf library yet.
Nice work! I see the code base is in tensorflow/tpu rather than in tensorflow/models or tensorflow/hub. Is it straightforward to run it on GPUs?
Thanks! Yes, it should run on GPU and TPU.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
